Lindblad-Inspired Multi-Timescale Reservoir Computing with Separable Rotation and Dissipation
arXiv:2608.04028
2026
Architecture
2 ideas extracted · analyzed Aug 31, 2026
What the math gives to ML
The paper provides a constructive recurrent parameterization in which each two-dimensional mode has an exactly controlled rotation angle and decay rate, rather than using an unconstrained random recurrent matrix followed by spectral-radius rescaling. Orthogonal mixing preserves the singular-value structure of the block-diagonal dynamics, so the largest decay factor gives an explicit global stability margin and the decay spectrum gives a bank of interpretable memory timescales. This structure can transfer directly to fixed reservoirs, recurrent layers, and linear state-space components as a constrained alternative to arbitrary recurrent matrices. The most promising experiments independently sweep rotation diversity and decay timescales at equal state size and compute budget.
Ideas from this paper
△ Mechanism confirmed, baseline not beaten
2026
Use a bank of damped rotational state channels with a deliberately spread decay spectrum, allowing one recurrent layer to represent short, medium, and long temporal dependencies without relying on a single learned spectral radius. Concatenate the channels and train a readout or downstream nonlinear head to select the appropriate memory timescale.
Useful7/10
Difficulty4/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Replace an unconstrained recurrent matrix with an orthogonally mixed block diagonal matrix whose blocks are independently parameterized damped rotations. The model receives explicit phase mixing from the rotation frequencies and controlled forgetting from the decay rates, while its linear recurrent dynamics have a known contraction factor before the nonlinear activation.
Useful7/10
Difficulty5/10
Novelty6/10